• Journal of Positioning, Navigation, and Timing
• /
• 제11권3호
• /
• pp.163-172
• /
• 2022

Typical anti-jamming technologies based on array antennas, Space Time Adaptive Process (STAP) & Space Frequency Adaptive Process (SFAP), are very effective algorithms to perform nulling and beamforming. However, it does not perform equally well for all types of jamming signals. If the anti-jamming algorithm is not optimized for each signal type, anti-jamming performance deteriorates and the operation stability of the system become worse by unnecessary computation. Therefore, jamming classification technique is required to obtain optimal anti-jamming performance. Machine learning, which has recently been in the spotlight, can be considered to classify jamming signal. In general, performing supervised learning for classification requires a huge amount of data and new learning for unfamiliar signal. In the case of jamming signal classification, it is difficult to obtain large amount of data because outdoor jamming signal reception environment is difficult to configure and the signal type of attacker is unknown. Therefore, this paper proposes few-shot jamming signal classification technique using meta-learning and transfer-learning to train the model using a small amount of data. A training dataset is constructed by anti-jamming algorithm input data within the GNSS receiver when jamming signals are applied. For meta-learning, Model-Agnostic Meta-Learning (MAML) algorithm with a general Convolution Neural Networks (CNN) model is used, and the same CNN model is used for transfer-learning. They are trained through episodic training using training datasets on developed our Python-based simulator. The results show both algorithms can be trained with less data and immediately respond to new signal types. Also, the performances of two algorithms are compared to determine which algorithm is more suitable for classifying jamming signals.

• 문화기술의 융합
• /
• 제8권1호
• /
• pp.545-550
• /
• 2022

딥러닝은 많은 데이터를 필요로 한다는 것은 이미 널리 알려져있다. 이를 통해, 딥러닝에 쓰이는 신경망의 수없이 많은 parameter들을 학습시킨다. 학습과정에는 데이터뿐 아니라, 각 데이터별로 전문가가 입력한 label이 필요한 경우가 대부분인데, 이 label을 얻는 과정은 시간과 자원 소비가 심하다. 이 문제를 완화하기 위해, few-shot learning, self-supervised learning, weak-supervised learning등이 연구되어오고 있다. 본 논문에서는, label을 상대적으로 적은 노력으로 수행하기 위한 연구들의 동향을 살펴보고, 앞으로의 개선 방향을 제시하도록 한다.

• 한국정보통신학회:학술대회논문집
• /
• 한국정보통신학회 2021년도 춘계학술대회
• /
• pp.332-333
• /
• 2021

본 논문에서는 소규모 데이터 세트의 이미지 분류 작업에서 모델 과적 합 및 비 수렴을 해결하고 분류의 정확성을 높이는 데 주로 사용되는 몇 번의 학습을 기반으로 한 새로운 이미지 분류 방법을 제안한다.

• Smart Structures and Systems
• /
• 제30권6호
• /
• pp.687-701
• /
• 2022

Inevitable response loss under complex operational conditions significantly affects the integrity and quality of measured data, leading the structural health monitoring (SHM) ineffective. To remedy the impact of data loss, a common way is to transfer the recorded response of available measure point to where the data loss occurred by establishing the response mapping from measured data. However, the current research has yet addressed the structural condition changes afterward and response mapping learning from a small sample. So, this paper proposes a novel data driven structural response reconstruction method based on a sophisticated designed generating adversarial network (UAGAN). Advanced deep learning techniques including U-shaped dense blocks, self-attention and a customized loss function are specialized and embedded in UAGAN to improve the universal and representative features extraction and generalized responses mapping establishment. In numerical validation, UAGAN efficiently and accurately captures the distinguished features of structural response from only 40 training samples of the intact structure. Besides, the established response mapping is universal, which effectively reconstructs responses of the structure suffered up to 10% random stiffness reduction or structural damage. In the experimental validation, UAGAN is trained with ambient response and applied to reconstruct response measured under earthquake. The reconstruction losses of response in the time and frequency domains reached 16% and 17%, that is better than the previous research, demonstrating the leading performance of the sophisticated designed network. In addition, the identified modal parameters from reconstructed and the corresponding true responses are highly consistent indicates that the proposed UAGAN is very potential to be applied to practical civil engineering.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제16권5호
• /
• pp.1755-1777
• /
• 2022

The development of wireless communication technology has led to the underutilization of radio spectra. To address this limitation, an intelligent cognitive radio network was developed. Specific emitter identification (SEI) is a key technology in this network. However, in realistic non-cooperative scenarios, the system may detect signal classes beyond those in the training database, and only a few labeled signal samples are available for network training, both of which deteriorate identification performance. To overcome these challenges, a meta-learning-based open-set identification system is proposed for SEI. First, the received signals were pre-processed using bi-spectral analysis and a Radon transform to obtain signal representation vectors, which were then fed into an open-set SEI network. This network consisted of a deep feature extractor and an intrinsic feature memorizer that can detect signals of unknown classes and classify signals of different known classes. The training loss functions and the procedures of the open-set SEI network were then designed for parameter optimization. Considering the few-shot problems of open-set SEI, meta-training loss functions and meta-training procedures that require only a few labeled signal samples were further developed for open-set SEI network training. The experimental results demonstrate that this approach outperforms other state-of-the-art SEI methods in open-set scenarios. In addition, excellent open-set SEI performance was achieved using at least 50 training signal samples, and effective operation in low signal-to-noise ratio (SNR) environments was demonstrated.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권10호
• /
• pp.3608-3626
• /
• 2021

The detection of bolts is an important task in high-speed train inspection systems, and it is frequently performed to ensure the safety of trains. The difficulty of the vision-based bolt inspection system lies in small sample defect detection, which makes the end-to-end network ineffective. In this paper, the problem is resolved in two stages, which includes the detection network and cascaded classification networks. For small bolt detection, all bolts including defective bolts and normal bolts are put together for conducting annotation training, a new loss function and a new boundingbox selection based on the smallest axis-aligned convex set are proposed. These allow YOLOv3 network to obtain the accurate position and bounding box of the various bolts. The average precision has been greatly improved on PASCAL VOC, MS COCO and actual data set. After that, the Siamese network is employed for estimating the status of the bolts. Using the convolutional Siamese network, we are able to get strong results on few-shot classification. Extensive experiments and comparisons on actual data set show that the system outperforms state-of-the-art algorithms in bolt inspection.

• 한국정보과학회 언어공학연구회:학술대회논문집(한글 및 한국어 정보처리)
• /
• 한국정보과학회언어공학연구회 2023년도 제35회 한글 및 한국어 정보처리 학술대회
• /
• pp.664-671
• /
• 2023

최근 다양한 거대 언어모델(Large Language Model)들이 개발되면서 프롬프트 엔지니어링의 대한 다양한 연구가 진행되고 있다. 본 논문에서는 퓨삿 학습 환경에서 개체명 인식의 성능을 높이기 위해서 제안된 템플릿이 필요 없는 프롬프트 튜닝(Template-free Prompt Tuning) 방법을 이용하고, 이 방법에서 사용된 라벨 단어 집합 생성 방법에 Maximal Marginal Relevance 알고리즘을 적용하여 해당 개체명에 대해 보다 다양하고 구체적인 라벨 단어 집합을 생성하도록 개선하였다. 실험 결과, 'LOC' 타입을 제외한 나머지 개체명 타입에서 'PER' 타입은 0.60%p, 'ORG' 타입은 4.98%p, 'MISC' 타입은 1.38%p 성능이 향상되었고, 전체 개체명 인식 성능은 1.26%p 향상되었다. 이를 통해 본 논문에서 제안한 라벨 단어 집합 생성 기법이 개체명 인식 성능 향상에 도움이 됨을 보였다.

• 정보처리학회 논문지
• /
• 제13권3호
• /
• pp.148-155
• /
• 2024

자연어 처리 분야에서 반어 및 비꼼 탐지의 중요성이 커지고 있음에도 불구하고, 한국어에 관한 연구는 다른 언어들에 비해 상대적으로 많이 부족한 편이다. 본 연구는 한국어 텍스트에서의 반어 탐지를 위해 다양한 모델을 실험하는 것을 목적으로 한다. 본 연구는 BERT기반 모델인 KoBERT와 ChatGPT를 사용하여 반어 탐지 실험을 수행하였다. KoBERT의 경우, 감성 데이터를 추가 학습하는 두 가지 방법(전이 학습, 멀티태스크 학습)을 적용하였다. 또한 ChatGPT의 경우, Few-Shot Learning기법을 적용하여 프롬프트에 입력되는 예시 문장의 개수를 증가시켜 실험하였다. 실험을 수행한 결과, 감성 데이터를 추가학습한 전이 학습 모델과 멀티태스크 학습 모델이 감성 데이터를 추가 학습하지 않은 기본 모델보다 우수한 성능을 보였다. 한편, ChatGPT는 KoBERT에 비해 현저히 낮은 성능을 나타내었으며, 입력 예시 문장의 개수를 증가시켜도 뚜렷한 성능 향상이 이루어지지 않았다. 종합적으로, 본 연구는 KoBERT를 기반으로 한 모델이 ChatGPT보다 반어 탐지에 더 적합하다는 결론을 도출했으며, 감성 데이터의 추가학습이 반어 탐지 성능 향상에 기여할 수 있는 가능성을 제시하였다.

• 한국융합학회논문지
• /
• 제12권11호
• /
• pp.35-43
• /
• 2021

최근 다언어모델(Cross-lingual language model)을 활용하여 한 번도 보지 못한 특정 언어의 하위 태스크를 수행하는 제로샷 교차언어 전이(Zero-shot cross-lingual transfer)에 대한 관심이 증가하고 있다. 본 논문은 기계번역 품질 예측(Quality Estimation, QE)을 학습하기 위한 데이터 구축적 측면에서의 한계점을 지적하고, 데이터를 구축하기 어려운 상황에서도 QE를 수행할 수 있도록 제로샷 교차언어 전이를 수행한다. QE에서 제로샷을 다룬 연구는 드물며, 본 논문에서는 교차언어모델을 활용하여 영어-독일어 QE 데이터에 대해 미세조정을 실시한 후 다른 언어쌍으로의 제로샷 전이를 진행했고 이 과정에서 다양한 다언어모델을 활용하여 비교 연구를 수행했다. 또한 다양한 자원 크기로 구성된 언어쌍에 대해 제로샷 실험을 진행하고 실험 결과에 대해 언어별 언어학적 특성 관점으로의 분석을 수행하였다. 실험결과 multilingual BART와 multillingual BERT에서 가장 높은 성능을 보였으며, 특정 언어쌍에 대해 QE 학습을 전혀 진행하지 않은 상황에서도 QE를 수행할 수 있도록 유도하였다.

• 천문학회보
• /
• 제44권1호
• /
• pp.54.2-54.2
• /
• 2019

Removing noise which occurs inevitably when taking image data has been a big concern. There is a way to raise signal-to-noise ratio and it is regarded as the only way, image stacking. Image stacking is averaging or just adding all pixel values of multiple pictures taken of a specific area. Its performance and reliability are unquestioned, but its weaknesses are also evident. Object with fast proper motion can be vanished, and most of all, it takes too long time. So if we can handle single shot image well and achieve similar performance, we can overcome those weaknesses. Recent developments in deep learning have enabled things that were not possible with former algorithm-based programming. One of the things is generating data with more information from data with less information. As a part of that, we reproduced stacked image from single shot image using a kind of deep learning, conditional generative adversarial network (cGAN). r-band camcol2 south data were used from SDSS Stripe 82 data. From all fields, image data which is stacked with only 22 individual images and, as a pair of stacked image, single pass data which were included in all stacked image were used. All used fields are cut in $128{\times}128$ pixel size, so total number of image is 17930. 14234 pairs of all images were used for training cGAN and 3696 pairs were used for verify the result. As a result, RMS error of pixel values between generated data from the best condition and target data were $7.67{\times}10^{-4}$ compared to original input data, $1.24{\times}10^{-3}$. We also applied to a few test galaxy images and generated images were similar to stacked images qualitatively compared to other de-noising methods. In addition, with photometry, The number count of stacked-cGAN matched sources is larger than that of single pass-stacked one, especially for fainter objects. Also, magnitude completeness became better in fainter objects. With this work, it is possible to observe reliably 1 magnitude fainter object.